A Feature Pyramid Network Based Detection Approach for Multi-Scale Target Detection in Stationary Short-Range Radar

This paper presents a deep learning-based Feature Pyramid Network (FPN) for object detection on range-Doppler maps in a stationary short-range radar setup operating at 60 GHz with 4 GHz bandwidth. The FPN model enables multiscale feature extraction, detecting large and small objects in complex, cluttered environments. It offers an alternative to traditional CA- and OS-CFAR detectors combined with DBSCAN, which can struggle with such variations. Holistic image processing through convolutional kernels in a one-stage architecture can help overcome limitations from clutter and noise points by enabling detection and clustering within a single process. Achieving an F1 score of 0.8683 at a 95% confidence threshold, the FPN balances precision (0.9050) and recall (0.8345), demonstrating competitiveness with OS-CFAR (F1: 0.8534) and CA-CFAR (F1: 0.8776). By capturing hierarchical features, the FPN learns contextual patterns beyond fixed-threshold methods, ensuring reliable detection of diverse target dimensions. Its real-time, edge-based data processing capabilities open opportunities for applications in smart homes, industrial automation, and infrastructure management.